Anthropic has released Claude Opus 5, the latest iteration in its Opus model lineup, positioning it as a highly capable option that approaches the performance of the company’s frontier Fable 5 model while remaining more accessible in terms of price and everyday usability.
Announced on July 24, 2026, the new model is now available and serves as the default for Claude Max subscribers while representing the top-tier choice for Claude Pro users.Opus 5 builds substantially on its predecessor, Opus 4.8, delivering stronger results across software engineering, knowledge work, and long-horizon agentic tasks without increasing token costs.
The company emphasizes that the model retains the same pricing structure as Opus 4.8—$5 per million input tokens and $25 per million output tokens—making it roughly half the cost of Fable 5 while closing much of the intelligence gap.
Users can adjust effort settings to balance higher intelligence against speed and token efficiency, allowing optimization for different workloads.
In key benchmarks, Opus 5 establishes new marks in several areas.
On Frontier-Bench, it more than doubles the performance of Opus 4.8 at lower cost per task and surpasses other competing models.
On CursorBench 3.2, it comes within half a percentage point of Fable 5’s peak score at maximum effort, yet achieves superior performance-per-cost ratios across high, extra-high, and maximum effort levels.
Additional gains appear on evaluations such as ARC-AGI 3, where scores reach three times those of the next-best model; Zapier AutomationBench, with roughly 1.5 times the pass rate of alternatives at equivalent cost; and OSWorld 2.0 for computer-use tasks, where it exceeds Fable 5’s best results at about one-third the cost.
Beyond coding, the model shows improvements in scientific and analytical domains.
It outperforms Opus 4.8 across internal life-sciences evaluations covering structural biology, organic chemistry, and bioinformatics, with particularly notable lifts in molecular structure inference from spectroscopy data and protein function prediction from sequence variations.
Visual generation capabilities have also advanced, enabling more sophisticated interactive illustrations and simulations.Early-access feedback highlights Opus 5’s stronger agency, self-verification, and thoroughness.
Testers reported instances where the model independently developed computer-vision pipelines to interpret incomplete inputs, identified root causes of bugs that other systems missed, constructed complete market-data feeds with custom validation harnesses, and maintained consistency across complex multi-step workflows.
Partners noted gains in debugging, root-cause analysis, genomics research, financial modeling, legal analysis, and long-horizon project execution, often with greater efficiency and fewer tokens or turns required.
On the safety front, Anthropic reports that Opus 5 is its most aligned model to date according to automated behavioral audits.
It demonstrates lower rates of deceptive behavior, stronger adherence to the company’s constitution, reduced susceptibility to misuse, and greater caution around irreversible actions.
While general capability improvements have enhanced its ability to identify software vulnerabilities, the model remains substantially behind Mythos-class systems in exploit development and does not push the frontier in high-risk dual-use domains such as advanced biology research or offensive cybersecurity.
Safeguards remain in place, consistent with prior Opus releases.
Overall, Claude Opus 5 aims to bring near-frontier intelligence into routine professional and developer workflows.
By combining elevated performance in coding, agentic reasoning, and knowledge work with cost efficiency and refined alignment, Anthropic positions the model as a practical daily driver rather than a specialized high-end tool. Availability extends across Claude interfaces and the API under the identifier claude-opus-5.
